Students increasingly build and use AI tools, but it is hard for them to judge how reliable those tools really are. Building an assistant is one skill, while testing where it breaks and explaining why is another, and that second skill matters most once AI becomes part of everyday work.

In this masters elective, students build their own AI assistant and then evaluate the assistants of two classmates. They rate each one on three areas: interaction and user experience, accuracy and reliability, and safety and risk management.

Students are asked to actively try to make the assistant fail, by finding edge cases and pushing it into unexpected territory. Reviews are anonymous, and grades are based on the quality of the written feedback students give, not on the scores they hand out.

In the final phase, students read the feedback on their own assistant, build a second version and reflect on what they changed, what they kept and why. As Michelle Sisto puts it: “The peer feedback is not an evaluation mechanism. It is a learning mechanism.”

Finding the flaws in a classmate's assistant, and then the same flaws in their own, gave students a grounded sense of what AI can and cannot do. With 30 students, one round produced more than 360 peer responses, and after three runs the design is now used in a course of 800 students.